Do the Universe's Empty Spaces Look the Way We Predict?

A clean null result from comparing observed and simulated cosmic voids.

Joe Poehlein

---

Update, July 2026: this paper was temporarily withdrawn along with the rest of this site's research corpus while every piece underwent a full re-audit against upgraded pipeline standards. The core result survived: every published statistic reproduces exactly from the archived pipeline, and the null holds under the corrected checks. This revision fixes two inaccurate statements the audit found the sample-selection step was misattributed to the source catalog, and the ellipticity caveat named the wrong confound and adds a selection-sensitivity analysis. Details in the Revision Note at the end.

---

The Question

The standard model of cosmology (ΛCDM) predicts how matter clumps and where it doesn't. The places where it doesn't, cosmic voids, are the largest structures in the universe: regions tens to hundreds of megaparsecs across, nearly empty of galaxies. The Boötes Void alone spans roughly 330 million light-years.

If ΛCDM is wrong, the size distribution of these voids should differ from what we observe. The Dark Energy Spectroscopic Instrument (DESI), the most ambitious galaxy survey to date, has now cataloged enough voids to test this.

We asked: do the void radius distributions from DESI observations match ΛCDM predictions from N-body simulation? We found no statistically significant difference in the radius distributions under three non-parametric tests.

Data

Observed voids. We used the DESIVAST DR1 catalogs (Rincon et al. 2025), a set of value-added void catalogs built from the DESI Data Release 1 Bright Galaxy Survey. DESIVAST publishes three separate void catalogs using different algorithms: VoidFinder, and V² with two pruning schemes (REVOLVER and VIDE). For this comparison we used the V² catalog with VIDE pruning (Sutter et al. 2015), which identifies voids out to redshift z ≤ 0.24.

Simulated voids. We ran VIDE on the AbacusSummit N-body simulation suite (Maksimova et al. 2021), specifically the small_c000_ph3000 box at redshift z = 0.2, using periodic boundary conditions. AbacusSummit uses Planck 2018 ΛCDM cosmology as its fiducial model. The simulations were produced using the Abacus N-body code (Garrison et al. 2021).

Both catalogs were processed with the same void-finding algorithm (VIDE), reducing algorithmic differences as a confound. However, using the same algorithm does not eliminate all confounds: differences in survey geometry, boundary effects, tracer selection, number density, and mask treatment between the observed survey and the periodic simulation box can all affect inferred void statistics. DESIVAST is built from a volume-limited BGS sample, which partially mitigates tracer-selection issues. The redshift mismatch (z ≤ 0.24 observed vs. z = 0.2 simulated) was resolved by downloading the z = 0.2 AbacusSummit snapshot rather than interpolating.

Sample selection. Our pipeline not the DESIVAST catalog itself, as an earlier version of this paper incorrectly stated excluded boundary-affected observed voids: any void with more than 25% of its surface area on the survey edge (EDGE_AREA / TOT_AREA > 0.25) was dropped. This removed 478 of 1,478 DESIVAST voids (32%), leaving the n = 1,000 analyzed below. On the simulation side, 4 of 4,854 voids with radii above a quarter of the box size were removed as periodic-boundary artifacts, leaving n = 4,850. Boundary-truncated voids have systematically underestimated radii, which is why interior selection is standard practice; the sensitivity of the result to this cut is reported in Method.

Method

Three non-parametric statistical tests on the void effective radius distributions:

1. Mann-Whitney U test tests whether one distribution is stochastically larger than the other. U = 2,502,235 (n_DESIVAST = 1,000, n_AbacusSummit = 4,850), p = 0.112, rank-biserial r = −0.032. No significant difference at α = 0.05.

2. Kolmogorov-Smirnov test tests whether the two samples are drawn from the same continuous distribution. D = 0.039, p = 0.148. No significant difference.

3. Permutation test 10,000 permutations of the combined radius samples, computing the difference in medians under the null hypothesis that the labels (observed vs. simulated) are exchangeable. Observed median difference = 0.190 Mpc/h, p = 0.385. The observed statistic falls well within the null distribution.

All three tests agree. There is no statistically significant difference between the observed and simulated void radius distributions.

Selection sensitivity. The 2026-07 re-audit re-ran all three tests across edge-fraction thresholds. The null is stable for interior-void selections (thresholds 0.05, 0.10, and the published 0.25 all give p > 0.05 on every test). With no edge filter at all including the 478 boundary-truncated voids all three tests flip significant (Mann-Whitney p = 0.032, KS p = 0.028, permutation p = 0.039). We attribute the flip to survey-boundary truncation biasing edge-void radii low rather than to cosmology, which is the standard reason for excluding boundary voids; but readers should know the null is conditional on interior selection.

Result

Clean null. No detectable radius discrepancy.

We found no statistically significant difference between the observed and simulated void radius distributions under three independent non-parametric tests. This clean null supports consistency between DESI low-redshift void radii and ΛCDM expectations from AbacusSummit under the present comparison.

The claim is scoped to what was tested: voids as defined by the V² algorithm with VIDE pruning, interior to the survey, at z ≲ 0.24, against a single simulation box. DESIVAST's other two population definitions (VoidFinder, and V² with REVOLVER pruning) define substantially different void populations and were not tested here; a same-algorithm comparison for each would be required to generalize the null beyond the VIDE-defined population.

Caveats

Three limitations should be noted honestly.

Ellipticity comparison was blocked for a different reason than we first reported. We attempted to compare void shapes (ellipticities) between observed and simulated catalogs. An earlier version of this paper attributed the problem to redshift-space distortions (RSD). The 2026-07 re-audit found the actual blocker is more basic: the shape-axis columns in the DESIVAST value-added catalog do not carry the same semantics as the raw VAST V² output we generated for the simulation. Applying the same axis-length formula to both yields observed "axes" whose median lengths (5.0, 27.7, 27.7 Mpc/h) are inconsistent with the observed median void radius (15.6 Mpc/h) two axes larger than the void itself while the simulated axes (15–16 Mpc/h) are self-consistent. The resulting distributions do not even overlap, which no physical effect, RSD included, can explain. The comparison remains deferred; doing it properly requires resolving the DESIVAST column conventions (and then still handling RSD, which is a real but secondary confound).

Single simulation box. We used one AbacusSummit box. Cosmic variance the statistical variation from box to box could shift the comparison. AbacusSummit provides ~2,000 small boxes; extending to a multi-box analysis would quantify this uncertainty. For this initial radius distribution comparison, the single-box result is sufficient to show no obvious tension, but tighter uncertainty bounds would require the full suite.

No density profile comparison. We compared void sizes but not void density profiles (the radial density run from void center to void wall). Density profiles are a more stringent test of the void model and are available in both catalogs. This is a natural Phase 2 extension.

Pipeline

The full analysis ran in approximately 8 minutes wall clock on a consumer workstation:

• ~5 minutes: data downloads (DESIVAST catalog + AbacusSummit snapshot)
• ~3.2 minutes: VIDE tessellation and void-finding (189 s, dominated by the serial tessellation stage)
• Negligible: statistical comparison (~3.4 seconds for 10,000 permutations, pure NumPy)

Three Python scripts: 00_prepare_abacus.py (download and format the simulation snapshot), 01_run_vide.py (run VIDE on both catalogs), 02_compare.py (statistical tests and output).

What This Means

This is a negative result in the best sense. We set up a test that could have found a discrepancy between observation and theory. It didn't. For VIDE-defined voids at low redshift, ΛCDM's predictions for the universe's large empty regions are holding up against a large new DESI-based catalog.

Negative results matter. They constrain the space of possible models by ruling out deviations that aren't there. They also establish a baseline: when future DESI data releases extend to higher redshifts and larger void populations, any emerging discrepancy will be measured against this null.

---

Revision Note

July 2026. After methodological flaws surfaced in the earliest project in this research program (the tornado-ERA5 work, since rewritten), every published piece on this site was withdrawn pending a full re-audit. This paper's audit re-derived every published statistic from the archived pipeline and raw catalogs (all reproduced exactly), re-ran the analysis across selection thresholds, and probed the alternative catalog definitions. The null result stands. Three corrections were made in this revision:

1. The boundary-void exclusion (EDGE_AREA/TOT_AREA ≤ 0.25, removing 32% of the observed sample) was performed by our pipeline, not by DESIVAST as previously stated, and is now disclosed with a sensitivity analysis showing the null is conditional on interior selection.
2. The ellipticity caveat previously blamed RSD; the actual blocker is a column-semantics mismatch between the DESIVAST catalog and raw VAST output.
3. The headline claim is now explicitly scoped to the V²/VIDE void population definition.

No numbers changed. The corrections are to what the paper said about its own method.

---

Author Note

The author is a software engineer, not a cosmologist. This analysis was conducted using an AI-assisted research pipeline: Claude (Anthropic) served as research coordinator and managed project state across sessions; Claude Code (Anthropic) executed the computational pipeline autonomously on consumer hardware. ChatGPT (OpenAI) provided independent adversarial review during technical review of the broader research arc this work belongs to.

All scientific decisions — hypothesis formation, experimental design, kill criteria, and interpretation — were made by the author. The AI tools were used as execution and review instruments, not as originators of the research questions or methodology.

---

References

Garrison, L. H., Eisenstein, D. J., Ferrer, D., Maksimova, N. A., & Pinto, P. A. (2021). The Abacus cosmological N-body code. Monthly Notices of the Royal Astronomical Society, 508(1), 575–596.

Maksimova, N. A., Garrison, L. H., Eisenstein, D. J., Hadzhiyska, B., Bose, S., & Satterthwaite, T. P. (2021). AbacusSummit: a massive set of high-accuracy, high-resolution N-body simulations. Monthly Notices of the Royal Astronomical Society, 508(3), 4017–4037.

Rincon, H. et al. (2025). DESIVAST: Catalogs of Low-redshift Voids Using Data from the DESI Data Release 1 Bright Galaxy Survey. The Astrophysical Journal, 982(1), 38. arXiv:2411.00148.

Sutter, P. M. et al. (2015). VIDE: The Void IDentification and Examination toolkit. Astronomy and Computing, 9, 1–9. arXiv:1406.1191.